WAN 2.2 I2V 14B + Lightning 2.2 I2V, Pusa, FusionX

screenshot of workflow being review by moderators currently...

Workflow that uses a bunch of helpers and doesn't crush the LoRa action / motion. Just threw it together and configured with various tips and bits of info from comments on the site. It works well. Can gen 480p or 720p.

8/7

• Uses the new Lightning 2.2 I2V, I was using a Char Lora, so the faces weren't an issue. If you are noticing issues with face variety, disable the high noise lightning lora. It gave me a tiny performance boost when enabled, but not really anything special. Might as well disable it.

  • but only on low noise. High noise Lightning is disabled re-enabled but see note above

  • After some more testing it looks like having them both enabled is the slightly better option see note above

• We also set the action LoRas to 2 on high noise and 1 on low noise. (sometimes, see Additional Info below)

Additional Info

With a powerful and strong LoRa like WAN Cumshot setting the high noise to 2.0 can affect prompt adherence negatively. If you set both high and low noise to 1.0 it will better adhere to prompt, but remember the trick if you need the boost on other LoRas.

Workflow set to 16fps but 24fps possible. With something like WAN cumshot LoRa enabled it needs to be set to 16 fps or it produces fast motion, my guess is it depends on what fps the lora was trained on. When using other LoRas (only tested a few) or just base model 24 fps works well. For example pov-tongue-job-wan-21-i2v works well at 24 fps and with this workflow

I'll continue to update this post with more information as I test and figure stuff out.

Contains

Fusion x phantom wan:

https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/blob/main/FusionX_LoRa/Phantom_Wan_14B_FusionX_LoRA.safetensors

Pusa:

https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Pusa/Wan21_PusaV1_LoRA_14B_rank512_bf16.safetensors

Lightning 2.2 I2V:

https://civarchive.com/models/1585622/self-forcing-causvid-accvid-lora-massive-speed-up-for-wan21-made-by-kijai?modelVersionId=2090344

Example LoRa used in this workflow:

https://civarchive.com/models/1350447/wan-cumshot-22-21